Deep Learning-Based Classification of NIFTP and Invasive Encapsulated Follicular Variant of Papillary Thyroid Carcinoma Using Gross Pathology Images.

Non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) and invasive encapsulated follicular variant of papillary thyroid carcinoma (IEFVPTC) are diagnostically challenging thyroid neoplasms with overlapping clinical and molecular characteristics. Although artificial intelligence has shown promise for diagnostic support in radiology and histopathology, its application to gross pathology remains unexplored. This study analyzed gross pathology photographs from 87 patients (43 with NIFTP and 44 with IEFVPTC). Three frozen, pretrained convolutional neural network backbones (EfficientNet, ResNet, and ConvNeXt) were used to generate feature embeddings, which were subsequently classified using Random Forest, support vector machine, and Gradient Boosting algorithms. Patient-level nested stratified five-fold cross-validation with inner hyperparameter tuning was used to generate unbiased out-of-fold predictions. EfficientNet combined with Random Forest was prespecified as the primary model and achieved a pooled area under the receiver operating characteristic curve (AUC) of 0.788 (95% CI, 0.614-0.928). Among the eight exploratory backbone-classifier combinations, pooled AUCs ranged from 0.500 to 0.755, and no pairwise comparison, including comparisons with the primary model, remained significantly different after Benjamini-Hochberg correction. The primary model demonstrated a sensitivity of 0.588, specificity of 0.944, accuracy of 0.771, and F1 score of 0.714. Gradient-weighted class activation mapping (Grad-CAM) revealed that all backbones predominantly highlighted lesional tissue rather than background regions, although activation patterns differed across architectures. In conclusion, this proof-of-concept study demonstrates that gross pathology images contain modest discriminative information for distinguishing NIFTP from IEFVPTC. These findings support the concept that macroscopic morphology encodes subtle yet class-separating signals.